Font Size: a A A

Research And Implementation Of Hybrid Neural Network Computing Platform Based On Hadoop

Posted on:2017-08-21Degree:MasterType:Thesis
Country:ChinaCandidate:L H LuoFull Text:PDF
GTID:2348330536953462Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
Artificial neural network aims to simulate the activities of the human brain using computer programs.Traditional shallow neural network or single type of neural network is usually only suitable for one kind of solution and not for the simulation study of the human brain.Deep neural network or hybrid neural network is made up by different types of neural networks,which are connected by means of series or parallel connections,are more capable of simulating the learning and memory process of human brain.The deep neural network or large-scale hybrid neural network requires a good deal of computing resources.The learning and training time of a large-scale hybrid neural network tend to be very long.Thus,a computing platform called HNetCP which is based on grid for hybrid neural networks was developed in this master thesis author's laboratory.Represented by Hadoop,cloud computing technology is one of the best choice to solve problems which require large-scale operation for most users.Based on the previous research results in the laboratory,MapReduce-based neural network on Hadoop is implemented in this master thesis,with the purpose of simulating the learning and memory process of human brain more efficiently.Face recognition is one of the most important application in simulating the learning and memory process of human brain.Dimension reduction and feature extraction of face recognition requires a good deal of computing resources.In the face of high dimensional images recognition problem,the efficiency of single computer is often too low to meet the realistic requirement.In consideration of the features of face recognition,using parallel principal component analysis and linear discriminant analysis to reducing dimensions and extracting features base on Map-Reduce framework is put forward in this master thesis.After that a RBF neural network classifier is implemented to classify the extracted features obtained in the previous stage.Serial algorithm and parallel algorithm are testing respectively on the ORL database of faces.The experiments showed that the parallel algorithm of feature extraction and network classification could improve the efficiency of face recognition to some extent.In addition,Deep learning aims to simulating the learning and memory process of human brain using multilayer neural network.Deep belief network is an important research area in the deep learning.The training efficiency of deep belief network becomes more important as the increasing data volume.A distributed deep belief network based on Hadoop is implemented in this master thesis.In the first stage,restricted boltzmann machines are trained in parallel to update the weight matrix and the bias of neurons.In the second stage,the whole network weight is fine tuning using BP algorithm.Testing on the MNIST database of handwriting digits showed that the training time of distributed deep belief network had a considerable reduction compared with serial algorithm.
Keywords/Search Tags:Hadoop, MapReduce-Based Neural Network, Face Recognition, Deep Belief Network
PDF Full Text Request
Related items